Home | History | Annotate | Line # | Download | only in libshmif
if_shmem.c revision 1.73
      1  1.73   msaitoh /*	$NetBSD: if_shmem.c,v 1.73 2017/10/23 09:31:18 msaitoh Exp $	*/
      2   1.1     pooka 
      3   1.1     pooka /*
      4  1.39     pooka  * Copyright (c) 2009, 2010 Antti Kantee.  All Rights Reserved.
      5   1.1     pooka  *
      6   1.1     pooka  * Development of this software was supported by The Nokia Foundation.
      7   1.1     pooka  *
      8   1.1     pooka  * Redistribution and use in source and binary forms, with or without
      9   1.1     pooka  * modification, are permitted provided that the following conditions
     10   1.1     pooka  * are met:
     11   1.1     pooka  * 1. Redistributions of source code must retain the above copyright
     12   1.1     pooka  *    notice, this list of conditions and the following disclaimer.
     13   1.1     pooka  * 2. Redistributions in binary form must reproduce the above copyright
     14   1.1     pooka  *    notice, this list of conditions and the following disclaimer in the
     15   1.1     pooka  *    documentation and/or other materials provided with the distribution.
     16   1.1     pooka  *
     17   1.1     pooka  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
     18   1.1     pooka  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     19   1.1     pooka  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
     20   1.1     pooka  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     21   1.1     pooka  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     22   1.1     pooka  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     23   1.1     pooka  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     24   1.1     pooka  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     25   1.1     pooka  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     26   1.1     pooka  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     27   1.1     pooka  * SUCH DAMAGE.
     28   1.1     pooka  */
     29   1.1     pooka 
     30   1.1     pooka #include <sys/cdefs.h>
     31  1.73   msaitoh __KERNEL_RCSID(0, "$NetBSD: if_shmem.c,v 1.73 2017/10/23 09:31:18 msaitoh Exp $");
     32   1.1     pooka 
     33   1.1     pooka #include <sys/param.h>
     34  1.13     pooka #include <sys/atomic.h>
     35   1.1     pooka #include <sys/fcntl.h>
     36   1.1     pooka #include <sys/kmem.h>
     37   1.1     pooka #include <sys/kthread.h>
     38   1.1     pooka #include <sys/lock.h>
     39  1.31     pooka #include <sys/vmem.h>
     40  1.44       tls #include <sys/cprng.h>
     41   1.1     pooka 
     42  1.35     pooka #include <net/bpf.h>
     43   1.1     pooka #include <net/if.h>
     44  1.37     pooka #include <net/if_dl.h>
     45   1.1     pooka #include <net/if_ether.h>
     46   1.1     pooka 
     47   1.1     pooka #include <netinet/in.h>
     48   1.1     pooka #include <netinet/in_var.h>
     49   1.1     pooka 
     50  1.64     pooka #include <rump-sys/kern.h>
     51  1.64     pooka #include <rump-sys/net.h>
     52  1.64     pooka 
     53   1.1     pooka #include <rump/rump.h>
     54   1.1     pooka #include <rump/rumpuser.h>
     55   1.1     pooka 
     56  1.59     pooka #include "shmif_user.h"
     57   1.1     pooka 
     58  1.29     pooka static int shmif_clone(struct if_clone *, int);
     59  1.29     pooka static int shmif_unclone(struct ifnet *);
     60  1.29     pooka 
     61  1.29     pooka struct if_clone shmif_cloner =
     62  1.29     pooka     IF_CLONE_INITIALIZER("shmif", shmif_clone, shmif_unclone);
     63  1.29     pooka 
     64   1.1     pooka /*
     65  1.28     pooka  * Do r/w prefault for backend pages when attaching the interface.
     66  1.29     pooka  * At least logically thinking improves performance (although no
     67  1.29     pooka  * mlocking is done, so they might go away).
     68  1.28     pooka  */
     69  1.28     pooka #define PREFAULT_RW
     70  1.28     pooka 
     71  1.28     pooka /*
     72   1.1     pooka  * A virtual ethernet interface which uses shared memory from a
     73   1.1     pooka  * memory mapped file as the bus.
     74   1.1     pooka  */
     75   1.1     pooka 
     76   1.1     pooka static int	shmif_init(struct ifnet *);
     77   1.1     pooka static int	shmif_ioctl(struct ifnet *, u_long, void *);
     78   1.1     pooka static void	shmif_start(struct ifnet *);
     79   1.1     pooka static void	shmif_stop(struct ifnet *, int);
     80   1.1     pooka 
     81  1.16     pooka #include "shmifvar.h"
     82  1.16     pooka 
     83   1.1     pooka struct shmif_sc {
     84   1.1     pooka 	struct ethercom sc_ec;
     85  1.16     pooka 	struct shmif_mem *sc_busmem;
     86   1.1     pooka 	int sc_memfd;
     87   1.1     pooka 	int sc_kq;
     88  1.32     pooka 	int sc_unit;
     89   1.1     pooka 
     90  1.29     pooka 	char *sc_backfile;
     91  1.29     pooka 	size_t sc_backfilelen;
     92  1.29     pooka 
     93  1.26     pooka 	uint64_t sc_devgen;
     94   1.1     pooka 	uint32_t sc_nextpacket;
     95  1.32     pooka 
     96  1.32     pooka 	kmutex_t sc_mtx;
     97  1.32     pooka 	kcondvar_t sc_cv;
     98  1.32     pooka 
     99  1.32     pooka 	struct lwp *sc_rcvl;
    100  1.32     pooka 	bool sc_dying;
    101  1.63     ozaki 
    102  1.63     ozaki 	uint64_t sc_uuid;
    103   1.1     pooka };
    104   1.1     pooka 
    105   1.1     pooka static void shmif_rcv(void *);
    106   1.1     pooka 
    107  1.23     pooka #define LOCK_UNLOCKED	0
    108  1.23     pooka #define LOCK_LOCKED	1
    109  1.23     pooka #define LOCK_COOLDOWN	1001
    110  1.23     pooka 
    111  1.31     pooka vmem_t *shmif_units;
    112  1.31     pooka 
    113  1.52     pooka static void
    114  1.52     pooka dowakeup(struct shmif_sc *sc)
    115  1.52     pooka {
    116  1.52     pooka 	struct rumpuser_iovec iov;
    117  1.52     pooka 	uint32_t ver = SHMIF_VERSION;
    118  1.53     pooka 	size_t n;
    119  1.52     pooka 
    120  1.52     pooka 	iov.iov_base = &ver;
    121  1.52     pooka 	iov.iov_len = sizeof(ver);
    122  1.53     pooka 	rumpuser_iovwrite(sc->sc_memfd, &iov, 1, IFMEM_WAKEUP, &n);
    123  1.52     pooka }
    124  1.52     pooka 
    125  1.23     pooka /*
    126  1.23     pooka  * This locking needs work and will misbehave severely if:
    127  1.23     pooka  * 1) the backing memory has to be paged in
    128  1.23     pooka  * 2) some lockholder exits while holding the lock
    129  1.23     pooka  */
    130  1.23     pooka static void
    131  1.23     pooka shmif_lockbus(struct shmif_mem *busmem)
    132  1.23     pooka {
    133  1.23     pooka 	int i = 0;
    134  1.23     pooka 
    135  1.23     pooka 	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
    136  1.23     pooka 	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
    137  1.23     pooka 		if (__predict_false(++i > LOCK_COOLDOWN)) {
    138  1.50     pooka 			/* wait 1ms */
    139  1.54     pooka 			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL,
    140  1.54     pooka 			    0, 1000*1000);
    141  1.23     pooka 			i = 0;
    142  1.23     pooka 		}
    143  1.23     pooka 		continue;
    144  1.23     pooka 	}
    145  1.23     pooka 	membar_enter();
    146  1.23     pooka }
    147  1.23     pooka 
    148  1.23     pooka static void
    149  1.23     pooka shmif_unlockbus(struct shmif_mem *busmem)
    150  1.23     pooka {
    151  1.61    justin 	unsigned int old __diagused;
    152  1.23     pooka 
    153  1.23     pooka 	membar_exit();
    154  1.23     pooka 	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
    155  1.23     pooka 	KASSERT(old == LOCK_LOCKED);
    156  1.23     pooka }
    157  1.23     pooka 
    158  1.29     pooka static int
    159  1.29     pooka allocif(int unit, struct shmif_sc **scp)
    160   1.1     pooka {
    161  1.29     pooka 	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
    162   1.1     pooka 	struct shmif_sc *sc;
    163   1.1     pooka 	struct ifnet *ifp;
    164   1.1     pooka 	uint32_t randnum;
    165  1.32     pooka 	int error;
    166   1.1     pooka 
    167  1.44       tls 	randnum = cprng_fast32();
    168  1.15     pooka 	memcpy(&enaddr[2], &randnum, sizeof(randnum));
    169   1.1     pooka 
    170   1.1     pooka 	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
    171  1.29     pooka 	sc->sc_memfd = -1;
    172  1.32     pooka 	sc->sc_unit = unit;
    173  1.63     ozaki 	sc->sc_uuid = cprng_fast64();
    174  1.29     pooka 
    175   1.1     pooka 	ifp = &sc->sc_ec.ec_if;
    176   1.1     pooka 
    177  1.60  christos 	snprintf(ifp->if_xname, sizeof(ifp->if_xname), "shmif%d", unit);
    178  1.29     pooka 	ifp->if_softc = sc;
    179  1.63     ozaki 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
    180  1.29     pooka 	ifp->if_init = shmif_init;
    181  1.29     pooka 	ifp->if_ioctl = shmif_ioctl;
    182  1.29     pooka 	ifp->if_start = shmif_start;
    183  1.29     pooka 	ifp->if_stop = shmif_stop;
    184  1.29     pooka 	ifp->if_mtu = ETHERMTU;
    185  1.37     pooka 	ifp->if_dlt = DLT_EN10MB;
    186  1.29     pooka 
    187  1.32     pooka 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
    188  1.32     pooka 	cv_init(&sc->sc_cv, "shmifcv");
    189  1.32     pooka 
    190  1.73   msaitoh 	error = if_initialize(ifp);
    191  1.73   msaitoh 	if (error != 0) {
    192  1.73   msaitoh 		aprint_error("shmif%d: if_initialize failed(%d)\n", unit, rv);
    193  1.73   msaitoh 		cv_destroy(&sc->sc_cv);
    194  1.73   msaitoh 		mutex_destroy(&sc->sc_mtx);
    195  1.73   msaitoh 		kmem_free(sc, sizeof(*sc));
    196  1.73   msaitoh 
    197  1.73   msaitoh 		return error;
    198  1.73   msaitoh 	}
    199  1.29     pooka 	ether_ifattach(ifp, enaddr);
    200  1.65     ozaki 	if_register(ifp);
    201  1.29     pooka 
    202  1.29     pooka 	aprint_verbose("shmif%d: Ethernet address %s\n",
    203  1.32     pooka 	    unit, ether_sprintf(enaddr));
    204  1.29     pooka 
    205  1.29     pooka 	if (scp)
    206  1.29     pooka 		*scp = sc;
    207  1.29     pooka 
    208  1.32     pooka 	error = 0;
    209  1.32     pooka 	if (rump_threads) {
    210  1.32     pooka 		error = kthread_create(PRI_NONE,
    211  1.40     rmind 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
    212  1.32     pooka 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
    213  1.32     pooka 	} else {
    214  1.32     pooka 		printf("WARNING: threads not enabled, shmif NOT working\n");
    215  1.32     pooka 	}
    216  1.32     pooka 
    217  1.32     pooka 	if (error) {
    218  1.32     pooka 		shmif_unclone(ifp);
    219  1.32     pooka 	}
    220  1.32     pooka 
    221  1.32     pooka 	return error;
    222  1.29     pooka }
    223  1.29     pooka 
    224  1.29     pooka static int
    225  1.29     pooka initbackend(struct shmif_sc *sc, int memfd)
    226  1.29     pooka {
    227  1.29     pooka 	volatile uint8_t v;
    228  1.29     pooka 	volatile uint8_t *p;
    229  1.53     pooka 	void *mem;
    230  1.29     pooka 	int error;
    231  1.29     pooka 
    232  1.53     pooka 	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
    233   1.1     pooka 	if (error)
    234  1.29     pooka 		return error;
    235  1.53     pooka 	sc->sc_busmem = mem;
    236  1.17     pooka 
    237  1.29     pooka 	if (sc->sc_busmem->shm_magic
    238  1.29     pooka 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
    239  1.29     pooka 		printf("bus is not magical");
    240  1.29     pooka 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    241  1.69   msaitoh 		return ENOEXEC;
    242  1.29     pooka 	}
    243  1.28     pooka 
    244  1.36     pooka 	/*
    245  1.36     pooka 	 * Prefault in pages to minimize runtime penalty with buslock.
    246  1.36     pooka 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
    247  1.36     pooka 	 * rump kernel PAGE_SIZE > host page size.
    248  1.36     pooka 	 */
    249  1.28     pooka 	for (p = (uint8_t *)sc->sc_busmem;
    250  1.28     pooka 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    251  1.36     pooka 	    p += 512)
    252  1.28     pooka 		v = *p;
    253  1.28     pooka 
    254  1.19     pooka 	shmif_lockbus(sc->sc_busmem);
    255  1.19     pooka 	/* we're first?  initialize bus */
    256  1.19     pooka 	if (sc->sc_busmem->shm_magic == 0) {
    257  1.19     pooka 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
    258  1.19     pooka 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
    259  1.19     pooka 	}
    260  1.19     pooka 
    261  1.16     pooka 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    262  1.26     pooka 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    263  1.28     pooka 
    264  1.28     pooka #ifdef PREFAULT_RW
    265  1.28     pooka 	for (p = (uint8_t *)sc->sc_busmem;
    266  1.28     pooka 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    267  1.28     pooka 	    p += PAGE_SIZE) {
    268  1.28     pooka 		v = *p;
    269  1.28     pooka 		*p = v;
    270  1.28     pooka 	}
    271  1.28     pooka #endif
    272  1.19     pooka 	shmif_unlockbus(sc->sc_busmem);
    273   1.1     pooka 
    274  1.53     pooka 	sc->sc_kq = -1;
    275  1.53     pooka 	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
    276  1.53     pooka 	if (error) {
    277  1.32     pooka 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    278  1.29     pooka 		return error;
    279  1.32     pooka 	}
    280   1.1     pooka 
    281  1.29     pooka 	sc->sc_memfd = memfd;
    282  1.32     pooka 
    283  1.32     pooka 	return error;
    284  1.29     pooka }
    285  1.29     pooka 
    286  1.29     pooka static void
    287  1.29     pooka finibackend(struct shmif_sc *sc)
    288  1.29     pooka {
    289  1.29     pooka 
    290  1.32     pooka 	if (sc->sc_backfile == NULL)
    291  1.32     pooka 		return;
    292  1.32     pooka 
    293  1.32     pooka 	if (sc->sc_backfile) {
    294  1.32     pooka 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
    295  1.32     pooka 		sc->sc_backfile = NULL;
    296  1.32     pooka 		sc->sc_backfilelen = 0;
    297  1.32     pooka 	}
    298  1.29     pooka 
    299  1.29     pooka 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    300  1.53     pooka 	rumpuser_close(sc->sc_memfd);
    301  1.53     pooka 	rumpuser_close(sc->sc_kq);
    302  1.32     pooka 
    303  1.32     pooka 	sc->sc_memfd = -1;
    304  1.29     pooka }
    305  1.29     pooka 
    306  1.29     pooka int
    307  1.29     pooka rump_shmif_create(const char *path, int *ifnum)
    308  1.29     pooka {
    309  1.29     pooka 	struct shmif_sc *sc;
    310  1.43    dyoung 	vmem_addr_t t;
    311  1.33     pooka 	int unit, error;
    312  1.33     pooka 	int memfd = -1; /* XXXgcc */
    313  1.29     pooka 
    314  1.33     pooka 	if (path) {
    315  1.53     pooka 		error = rumpuser_open(path,
    316  1.53     pooka 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    317  1.53     pooka 		if (error)
    318  1.33     pooka 			return error;
    319  1.33     pooka 	}
    320   1.1     pooka 
    321  1.43    dyoung 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
    322  1.43    dyoung 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
    323  1.43    dyoung 
    324  1.43    dyoung 	if (error != 0) {
    325  1.43    dyoung 		if (path)
    326  1.53     pooka 			rumpuser_close(memfd);
    327  1.43    dyoung 		return error;
    328  1.43    dyoung 	}
    329  1.43    dyoung 
    330  1.43    dyoung 	unit = t - 1;
    331  1.31     pooka 
    332  1.32     pooka 	if ((error = allocif(unit, &sc)) != 0) {
    333  1.33     pooka 		if (path)
    334  1.53     pooka 			rumpuser_close(memfd);
    335  1.29     pooka 		return error;
    336  1.29     pooka 	}
    337  1.33     pooka 
    338  1.33     pooka 	if (!path)
    339  1.33     pooka 		goto out;
    340  1.33     pooka 
    341  1.29     pooka 	error = initbackend(sc, memfd);
    342  1.29     pooka 	if (error) {
    343  1.32     pooka 		shmif_unclone(&sc->sc_ec.ec_if);
    344  1.29     pooka 		return error;
    345  1.29     pooka 	}
    346   1.1     pooka 
    347  1.29     pooka 	sc->sc_backfilelen = strlen(path)+1;
    348  1.29     pooka 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
    349  1.29     pooka 	strcpy(sc->sc_backfile, path);
    350  1.12     pooka 
    351  1.33     pooka  out:
    352   1.2     pooka 	if (ifnum)
    353  1.32     pooka 		*ifnum = unit;
    354  1.29     pooka 
    355   1.1     pooka 	return 0;
    356  1.29     pooka }
    357   1.1     pooka 
    358  1.29     pooka static int
    359  1.29     pooka shmif_clone(struct if_clone *ifc, int unit)
    360  1.29     pooka {
    361  1.61    justin 	int rc __diagused;
    362  1.43    dyoung 	vmem_addr_t unit2;
    363  1.29     pooka 
    364  1.31     pooka 	/*
    365  1.31     pooka 	 * Ok, we know the unit number, but we must still reserve it.
    366  1.31     pooka 	 * Otherwise the wildcard-side of things might get the same one.
    367  1.31     pooka 	 * This is slightly offset-happy due to vmem.  First, we offset
    368  1.31     pooka 	 * the range of unit numbers by +1 since vmem cannot deal with
    369  1.41    dyoung 	 * ranges starting from 0.  Talk about uuuh.
    370  1.31     pooka 	 */
    371  1.43    dyoung 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
    372  1.43    dyoung 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
    373  1.43    dyoung 	KASSERT(rc == 0 && unit2-1 == unit);
    374  1.29     pooka 
    375  1.29     pooka 	return allocif(unit, NULL);
    376  1.29     pooka }
    377  1.29     pooka 
    378  1.29     pooka static int
    379  1.29     pooka shmif_unclone(struct ifnet *ifp)
    380  1.29     pooka {
    381  1.32     pooka 	struct shmif_sc *sc = ifp->if_softc;
    382  1.32     pooka 
    383  1.32     pooka 	shmif_stop(ifp, 1);
    384  1.32     pooka 	if_down(ifp);
    385  1.32     pooka 
    386  1.32     pooka 	mutex_enter(&sc->sc_mtx);
    387  1.32     pooka 	sc->sc_dying = true;
    388  1.32     pooka 	cv_broadcast(&sc->sc_cv);
    389  1.32     pooka 	mutex_exit(&sc->sc_mtx);
    390  1.29     pooka 
    391  1.32     pooka 	if (sc->sc_rcvl)
    392  1.32     pooka 		kthread_join(sc->sc_rcvl);
    393  1.32     pooka 	sc->sc_rcvl = NULL;
    394  1.32     pooka 
    395  1.72     ozaki 	/*
    396  1.72     ozaki 	 * Need to be called after the kthread left, otherwise closing kqueue
    397  1.72     ozaki 	 * (sc_kq) hangs sometimes perhaps because of a race condition between
    398  1.72     ozaki 	 * close and kevent in the kthread on the kqueue.
    399  1.72     ozaki 	 */
    400  1.72     ozaki 	finibackend(sc);
    401  1.72     ozaki 
    402  1.32     pooka 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
    403  1.32     pooka 
    404  1.32     pooka 	ether_ifdetach(ifp);
    405  1.32     pooka 	if_detach(ifp);
    406  1.32     pooka 
    407  1.32     pooka 	cv_destroy(&sc->sc_cv);
    408  1.32     pooka 	mutex_destroy(&sc->sc_mtx);
    409  1.32     pooka 
    410  1.32     pooka 	kmem_free(sc, sizeof(*sc));
    411  1.32     pooka 
    412  1.32     pooka 	return 0;
    413   1.1     pooka }
    414   1.1     pooka 
    415   1.1     pooka static int
    416   1.1     pooka shmif_init(struct ifnet *ifp)
    417   1.1     pooka {
    418  1.29     pooka 	struct shmif_sc *sc = ifp->if_softc;
    419   1.4     pooka 	int error = 0;
    420   1.4     pooka 
    421  1.29     pooka 	if (sc->sc_memfd == -1)
    422  1.29     pooka 		return ENXIO;
    423  1.32     pooka 	KASSERT(sc->sc_busmem);
    424  1.29     pooka 
    425  1.32     pooka 	ifp->if_flags |= IFF_RUNNING;
    426  1.32     pooka 
    427  1.32     pooka 	mutex_enter(&sc->sc_mtx);
    428  1.32     pooka 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    429  1.32     pooka 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    430  1.32     pooka 
    431  1.32     pooka 	cv_broadcast(&sc->sc_cv);
    432  1.32     pooka 	mutex_exit(&sc->sc_mtx);
    433   1.1     pooka 
    434   1.4     pooka 	return error;
    435   1.1     pooka }
    436   1.1     pooka 
    437   1.1     pooka static int
    438   1.1     pooka shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
    439   1.1     pooka {
    440  1.29     pooka 	struct shmif_sc *sc = ifp->if_softc;
    441  1.29     pooka 	struct ifdrv *ifd;
    442  1.29     pooka 	char *path;
    443  1.32     pooka 	int s, rv, memfd;
    444   1.1     pooka 
    445   1.1     pooka 	s = splnet();
    446  1.29     pooka 	switch (cmd) {
    447  1.29     pooka 	case SIOCGLINKSTR:
    448  1.29     pooka 		ifd = data;
    449  1.29     pooka 
    450  1.29     pooka 		if (sc->sc_backfilelen == 0) {
    451  1.29     pooka 			rv = ENOENT;
    452  1.29     pooka 			break;
    453  1.29     pooka 		}
    454  1.29     pooka 
    455  1.29     pooka 		ifd->ifd_len = sc->sc_backfilelen;
    456  1.29     pooka 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
    457  1.29     pooka 			rv = 0;
    458  1.29     pooka 			break;
    459  1.29     pooka 		}
    460  1.29     pooka 
    461  1.29     pooka 		if (ifd->ifd_cmd != 0) {
    462  1.29     pooka 			rv = EINVAL;
    463  1.29     pooka 			break;
    464  1.29     pooka 		}
    465  1.29     pooka 
    466  1.29     pooka 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
    467  1.29     pooka 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
    468  1.29     pooka 		break;
    469  1.29     pooka 	case SIOCSLINKSTR:
    470  1.29     pooka 		if (ifp->if_flags & IFF_UP) {
    471  1.29     pooka 			rv = EBUSY;
    472  1.29     pooka 			break;
    473  1.29     pooka 		}
    474  1.29     pooka 
    475  1.29     pooka 		ifd = data;
    476  1.29     pooka 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
    477  1.29     pooka 			finibackend(sc);
    478  1.29     pooka 			rv = 0;
    479  1.29     pooka 			break;
    480  1.29     pooka 		} else if (ifd->ifd_cmd != 0) {
    481  1.29     pooka 			rv = EINVAL;
    482  1.29     pooka 			break;
    483  1.29     pooka 		} else if (sc->sc_backfile) {
    484  1.29     pooka 			rv = EBUSY;
    485  1.29     pooka 			break;
    486  1.29     pooka 		}
    487  1.29     pooka 
    488  1.29     pooka 		if (ifd->ifd_len > MAXPATHLEN) {
    489  1.29     pooka 			rv = E2BIG;
    490  1.29     pooka 			break;
    491  1.29     pooka 		} else if (ifd->ifd_len < 1) {
    492  1.29     pooka 			rv = EINVAL;
    493  1.29     pooka 			break;
    494  1.29     pooka 		}
    495  1.29     pooka 
    496  1.29     pooka 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
    497  1.29     pooka 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
    498  1.29     pooka 		if (rv) {
    499  1.29     pooka 			kmem_free(path, ifd->ifd_len);
    500  1.29     pooka 			break;
    501  1.29     pooka 		}
    502  1.53     pooka 		rv = rumpuser_open(path,
    503  1.53     pooka 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    504  1.53     pooka 		if (rv) {
    505  1.29     pooka 			kmem_free(path, ifd->ifd_len);
    506  1.29     pooka 			break;
    507  1.29     pooka 		}
    508  1.29     pooka 		rv = initbackend(sc, memfd);
    509  1.29     pooka 		if (rv) {
    510  1.29     pooka 			kmem_free(path, ifd->ifd_len);
    511  1.53     pooka 			rumpuser_close(memfd);
    512  1.29     pooka 			break;
    513  1.29     pooka 		}
    514  1.29     pooka 		sc->sc_backfile = path;
    515  1.29     pooka 		sc->sc_backfilelen = ifd->ifd_len;
    516  1.29     pooka 
    517  1.29     pooka 		break;
    518  1.29     pooka 	default:
    519  1.29     pooka 		rv = ether_ioctl(ifp, cmd, data);
    520  1.29     pooka 		if (rv == ENETRESET)
    521  1.29     pooka 			rv = 0;
    522  1.29     pooka 		break;
    523  1.29     pooka 	}
    524   1.1     pooka 	splx(s);
    525   1.1     pooka 
    526   1.1     pooka 	return rv;
    527   1.1     pooka }
    528   1.1     pooka 
    529  1.32     pooka /* send everything in-context since it's just a matter of mem-to-mem copy */
    530   1.1     pooka static void
    531   1.1     pooka shmif_start(struct ifnet *ifp)
    532   1.1     pooka {
    533   1.1     pooka 	struct shmif_sc *sc = ifp->if_softc;
    534  1.26     pooka 	struct shmif_mem *busmem = sc->sc_busmem;
    535   1.1     pooka 	struct mbuf *m, *m0;
    536  1.26     pooka 	uint32_t dataoff;
    537  1.26     pooka 	uint32_t pktsize, pktwrote;
    538   1.1     pooka 	bool wrote = false;
    539  1.24     pooka 	bool wrap;
    540   1.1     pooka 
    541  1.26     pooka 	ifp->if_flags |= IFF_OACTIVE;
    542  1.26     pooka 
    543   1.1     pooka 	for (;;) {
    544  1.20     pooka 		struct shmif_pkthdr sp;
    545  1.20     pooka 		struct timeval tv;
    546  1.20     pooka 
    547   1.1     pooka 		IF_DEQUEUE(&ifp->if_snd, m0);
    548   1.1     pooka 		if (m0 == NULL) {
    549   1.1     pooka 			break;
    550   1.1     pooka 		}
    551   1.1     pooka 
    552  1.25     pooka 		pktsize = 0;
    553  1.19     pooka 		for (m = m0; m != NULL; m = m->m_next) {
    554  1.19     pooka 			pktsize += m->m_len;
    555  1.19     pooka 		}
    556  1.26     pooka 		KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
    557  1.19     pooka 
    558  1.20     pooka 		getmicrouptime(&tv);
    559  1.20     pooka 		sp.sp_len = pktsize;
    560  1.20     pooka 		sp.sp_sec = tv.tv_sec;
    561  1.20     pooka 		sp.sp_usec = tv.tv_usec;
    562  1.63     ozaki 		sp.sp_sender = sc->sc_uuid;
    563  1.20     pooka 
    564  1.35     pooka 		bpf_mtap(ifp, m0);
    565  1.35     pooka 
    566  1.26     pooka 		shmif_lockbus(busmem);
    567  1.26     pooka 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    568  1.26     pooka 		busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
    569  1.21     pooka 
    570  1.24     pooka 		wrap = false;
    571  1.26     pooka 		dataoff = shmif_buswrite(busmem,
    572  1.26     pooka 		    busmem->shm_last, &sp, sizeof(sp), &wrap);
    573  1.26     pooka 		pktwrote = 0;
    574   1.1     pooka 		for (m = m0; m != NULL; m = m->m_next) {
    575  1.26     pooka 			pktwrote += m->m_len;
    576  1.26     pooka 			dataoff = shmif_buswrite(busmem, dataoff,
    577  1.19     pooka 			    mtod(m, void *), m->m_len, &wrap);
    578   1.1     pooka 		}
    579  1.26     pooka 		KASSERT(pktwrote == pktsize);
    580  1.27     pooka 		if (wrap) {
    581  1.26     pooka 			busmem->shm_gen++;
    582  1.47     pooka 			DPRINTF(("bus generation now %" PRIu64 "\n",
    583  1.47     pooka 			    busmem->shm_gen));
    584  1.27     pooka 		}
    585  1.26     pooka 		shmif_unlockbus(busmem);
    586   1.1     pooka 
    587   1.1     pooka 		m_freem(m0);
    588   1.1     pooka 		wrote = true;
    589  1.62     ozaki 		ifp->if_opackets++;
    590   1.1     pooka 
    591   1.1     pooka 		DPRINTF(("shmif_start: send %d bytes at off %d\n",
    592  1.27     pooka 		    pktsize, busmem->shm_last));
    593   1.1     pooka 	}
    594  1.26     pooka 
    595  1.26     pooka 	ifp->if_flags &= ~IFF_OACTIVE;
    596  1.26     pooka 
    597  1.32     pooka 	/* wakeup? */
    598  1.52     pooka 	if (wrote) {
    599  1.52     pooka 		dowakeup(sc);
    600  1.52     pooka 	}
    601   1.1     pooka }
    602   1.1     pooka 
    603   1.1     pooka static void
    604   1.1     pooka shmif_stop(struct ifnet *ifp, int disable)
    605   1.1     pooka {
    606  1.32     pooka 	struct shmif_sc *sc = ifp->if_softc;
    607   1.1     pooka 
    608  1.32     pooka 	ifp->if_flags &= ~IFF_RUNNING;
    609  1.32     pooka 	membar_producer();
    610  1.32     pooka 
    611  1.32     pooka 	/*
    612  1.32     pooka 	 * wakeup thread.  this will of course wake up all bus
    613  1.32     pooka 	 * listeners, but that's life.
    614  1.32     pooka 	 */
    615  1.52     pooka 	if (sc->sc_memfd != -1) {
    616  1.52     pooka 		dowakeup(sc);
    617  1.52     pooka 	}
    618   1.1     pooka }
    619   1.1     pooka 
    620  1.27     pooka 
    621  1.27     pooka /*
    622  1.27     pooka  * Check if we have been sleeping too long.  Basically,
    623  1.27     pooka  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
    624  1.27     pooka  * We use the fact that first is guaranteed to never overlap
    625  1.27     pooka  * with the last frame in the ring.
    626  1.27     pooka  */
    627  1.27     pooka static __inline bool
    628  1.27     pooka stillvalid_p(struct shmif_sc *sc)
    629  1.27     pooka {
    630  1.27     pooka 	struct shmif_mem *busmem = sc->sc_busmem;
    631  1.27     pooka 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
    632  1.27     pooka 	uint32_t lastoff, devoff;
    633  1.27     pooka 
    634  1.27     pooka 	KASSERT(busmem->shm_first != busmem->shm_last);
    635  1.27     pooka 
    636  1.27     pooka 	/* normalize onto a 2x busmem chunk */
    637  1.27     pooka 	devoff = sc->sc_nextpacket;
    638  1.27     pooka 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
    639  1.27     pooka 
    640  1.27     pooka 	/* trivial case */
    641  1.27     pooka 	if (gendiff > 1)
    642  1.27     pooka 		return false;
    643  1.27     pooka 	KASSERT(gendiff <= 1);
    644  1.27     pooka 
    645  1.27     pooka 	/* Normalize onto 2x busmem chunk */
    646  1.27     pooka 	if (busmem->shm_first >= lastoff) {
    647  1.27     pooka 		lastoff += BUSMEM_DATASIZE;
    648  1.27     pooka 		if (gendiff == 0)
    649  1.27     pooka 			devoff += BUSMEM_DATASIZE;
    650  1.27     pooka 	} else {
    651  1.27     pooka 		if (gendiff)
    652  1.27     pooka 			return false;
    653  1.27     pooka 	}
    654  1.27     pooka 
    655  1.27     pooka 	return devoff >= busmem->shm_first && devoff <= lastoff;
    656  1.27     pooka }
    657  1.27     pooka 
    658   1.1     pooka static void
    659   1.1     pooka shmif_rcv(void *arg)
    660   1.1     pooka {
    661   1.1     pooka 	struct ifnet *ifp = arg;
    662   1.1     pooka 	struct shmif_sc *sc = ifp->if_softc;
    663  1.32     pooka 	struct shmif_mem *busmem;
    664   1.1     pooka 	struct mbuf *m = NULL;
    665   1.1     pooka 	struct ether_header *eth;
    666  1.27     pooka 	uint32_t nextpkt;
    667  1.37     pooka 	bool wrap, passup;
    668   1.1     pooka 	int error;
    669  1.57     pooka 	const int align
    670  1.57     pooka 	    = ALIGN(sizeof(struct ether_header)) - sizeof(struct ether_header);
    671   1.1     pooka 
    672  1.32     pooka  reup:
    673  1.32     pooka 	mutex_enter(&sc->sc_mtx);
    674  1.32     pooka 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
    675  1.32     pooka 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
    676  1.32     pooka 	mutex_exit(&sc->sc_mtx);
    677  1.32     pooka 
    678  1.32     pooka 	busmem = sc->sc_busmem;
    679  1.32     pooka 
    680  1.32     pooka 	while (ifp->if_flags & IFF_RUNNING) {
    681  1.20     pooka 		struct shmif_pkthdr sp;
    682  1.20     pooka 
    683   1.1     pooka 		if (m == NULL) {
    684   1.1     pooka 			m = m_gethdr(M_WAIT, MT_DATA);
    685   1.1     pooka 			MCLGET(m, M_WAIT);
    686  1.57     pooka 			m->m_data += align;
    687   1.1     pooka 		}
    688   1.1     pooka 
    689  1.47     pooka 		DPRINTF(("waiting %d/%" PRIu64 "\n",
    690  1.47     pooka 		    sc->sc_nextpacket, sc->sc_devgen));
    691  1.26     pooka 		KASSERT(m->m_flags & M_EXT);
    692   1.1     pooka 
    693  1.26     pooka 		shmif_lockbus(busmem);
    694  1.26     pooka 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    695  1.27     pooka 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
    696   1.1     pooka 
    697   1.1     pooka 		/* need more data? */
    698  1.27     pooka 		if (sc->sc_devgen == busmem->shm_gen &&
    699  1.26     pooka 		    shmif_nextpktoff(busmem, busmem->shm_last)
    700  1.26     pooka 		     == sc->sc_nextpacket) {
    701  1.26     pooka 			shmif_unlockbus(busmem);
    702  1.71     ozaki 			error = rumpcomp_shmif_watchwait(sc->sc_kq);
    703   1.1     pooka 			if (__predict_false(error))
    704   1.1     pooka 				printf("shmif_rcv: wait failed %d\n", error);
    705  1.32     pooka 			membar_consumer();
    706   1.1     pooka 			continue;
    707   1.1     pooka 		}
    708   1.1     pooka 
    709  1.27     pooka 		if (stillvalid_p(sc)) {
    710  1.27     pooka 			nextpkt = sc->sc_nextpacket;
    711  1.27     pooka 		} else {
    712  1.27     pooka 			KASSERT(busmem->shm_gen > 0);
    713  1.26     pooka 			nextpkt = busmem->shm_first;
    714  1.26     pooka 			if (busmem->shm_first > busmem->shm_last)
    715  1.27     pooka 				sc->sc_devgen = busmem->shm_gen - 1;
    716  1.26     pooka 			else
    717  1.27     pooka 				sc->sc_devgen = busmem->shm_gen;
    718  1.47     pooka 			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
    719  1.27     pooka 			    sc, nextpkt, sc->sc_devgen));
    720  1.26     pooka 		}
    721  1.26     pooka 
    722  1.26     pooka 		/*
    723  1.26     pooka 		 * If our read pointer is ahead the bus last write, our
    724  1.26     pooka 		 * generation must be one behind.
    725  1.26     pooka 		 */
    726  1.26     pooka 		KASSERT(!(nextpkt > busmem->shm_last
    727  1.27     pooka 		    && sc->sc_devgen == busmem->shm_gen));
    728  1.26     pooka 
    729  1.24     pooka 		wrap = false;
    730  1.26     pooka 		nextpkt = shmif_busread(busmem, &sp,
    731  1.26     pooka 		    nextpkt, sizeof(sp), &wrap);
    732  1.26     pooka 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
    733  1.26     pooka 		nextpkt = shmif_busread(busmem, mtod(m, void *),
    734  1.26     pooka 		    nextpkt, sp.sp_len, &wrap);
    735   1.1     pooka 
    736   1.1     pooka 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
    737  1.20     pooka 		    sp.sp_len, nextpkt));
    738   1.1     pooka 
    739  1.26     pooka 		sc->sc_nextpacket = nextpkt;
    740  1.19     pooka 		shmif_unlockbus(sc->sc_busmem);
    741   1.1     pooka 
    742  1.27     pooka 		if (wrap) {
    743  1.26     pooka 			sc->sc_devgen++;
    744  1.47     pooka 			DPRINTF(("dev %p generation now %" PRIu64 "\n",
    745  1.27     pooka 			    sc, sc->sc_devgen));
    746  1.27     pooka 		}
    747  1.26     pooka 
    748  1.56     pooka 		/*
    749  1.56     pooka 		 * Ignore packets too short to possibly be valid.
    750  1.56     pooka 		 * This is hit at least for the first frame on a new bus.
    751  1.56     pooka 		 */
    752  1.55     pooka 		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
    753  1.55     pooka 			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
    754  1.55     pooka 			    sp.sp_len));
    755  1.55     pooka 			continue;
    756  1.55     pooka 		}
    757  1.55     pooka 
    758  1.20     pooka 		m->m_len = m->m_pkthdr.len = sp.sp_len;
    759  1.67     ozaki 		m_set_rcvif(m, ifp);
    760   1.1     pooka 
    761  1.37     pooka 		/*
    762  1.37     pooka 		 * Test if we want to pass the packet upwards
    763  1.37     pooka 		 */
    764   1.1     pooka 		eth = mtod(m, struct ether_header *);
    765  1.63     ozaki 		if (sp.sp_sender == sc->sc_uuid) {
    766  1.63     ozaki 			passup = false;
    767  1.63     ozaki 		} else if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
    768  1.37     pooka 		    ETHER_ADDR_LEN) == 0) {
    769  1.37     pooka 			passup = true;
    770  1.46     pooka 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
    771  1.37     pooka 			passup = true;
    772  1.37     pooka 		} else if (ifp->if_flags & IFF_PROMISC) {
    773  1.37     pooka 			m->m_flags |= M_PROMISC;
    774  1.37     pooka 			passup = true;
    775  1.38     pooka 		} else {
    776  1.38     pooka 			passup = false;
    777  1.37     pooka 		}
    778  1.37     pooka 
    779  1.37     pooka 		if (passup) {
    780  1.68     ozaki 			int bound;
    781  1.22     pooka 			KERNEL_LOCK(1, NULL);
    782  1.66     ozaki 			/* Prevent LWP migrations between CPUs for psref(9) */
    783  1.68     ozaki 			bound = curlwp_bind();
    784  1.65     ozaki 			if_input(ifp, m);
    785  1.68     ozaki 			curlwp_bindx(bound);
    786  1.22     pooka 			KERNEL_UNLOCK_ONE(NULL);
    787   1.1     pooka 			m = NULL;
    788   1.1     pooka 		}
    789  1.37     pooka 		/* else: reuse mbuf for a future packet */
    790   1.1     pooka 	}
    791  1.32     pooka 	m_freem(m);
    792  1.32     pooka 	m = NULL;
    793  1.32     pooka 
    794  1.32     pooka 	if (!sc->sc_dying)
    795  1.32     pooka 		goto reup;
    796   1.1     pooka 
    797  1.32     pooka 	kthread_exit(0);
    798   1.1     pooka }
    799